As we continue to experience the rapid evolution of AI technology, one thing is clear: AI used to augment human capabilities will yield the most impactful results.
Take, for example, the first study we share in this month’s digest. A doctor using AI was 2.6% better at detecting breast cancer than a doctor working without it. Moreover, these results were better than when the AI performed on its own.
Whether we’re working with models used for anomaly detection like this one, or language generation like ChatGPT, we must invest resources in the model-human interactions that will ultimately dictate their success rates.
How are you addressing this in your own organization?
P.S. Don’t forget to check out "Team Panda Picks" at the end for a list of articles curated by our team of data scientists.
What Do We Mean by Trusted AI?
Trusted AI is the discipline of designing AI-powered solutions that maximize the value of humans while being more fair, transparent, and privacy-preserving.
Doctors Use AI For More Successful Cancer Screening
What it's about: According to new research, doctors working with artificial intelligence screen for breast cancer more successfully than when working alone. Published in The Lancet Digital Health, this is the first study in breast cancer screening to compare how an AI performs on its own, to how it performs when used with a human expert. The software comes from a startup based in Germany called Vara, which also spearheaded the study. This specific AI has already been implemented in over a quarter of Germany’s breast cancer screening centers and has been introduced to hospitals in Mexico and Greece.
The team at Vara worked with radiologists at the Essen University Hospital in Germany and the Memorial Sloan Kettering Cancer Center in New York to test two approaches. The first approach studied how the AI worked alone to analyze mammograms.
In the second approach, the AI automatically differentiates between normal and abnormal scans. Then, the AI refers any concerning scans to a radiologist, who reviews them before seeing what the AI labeled them as. If the doctor did not detect cancer on a scan that the AI did, it issues a warning. In the study, the AI inspected historic scans. Its assessments were then compared with those of the original radiologist who reviewed the scans.
Why it matters: According to the article, the second approach of the doctor using AI was 2.6% better at detecting breast cancer and raised fewer false alarms than a doctor working without it. Additionally, the AI automatically distinguished scans as “confident normal,” which is 63% of all mammograms. This streamlined process could significantly reduce radiologists’ workload. Fatigue, overwork, and time of day can all affect how successfully radiologists can identify tumors. Today, radiologists examining scans miss one in eight cancers.
We’re entering a period where AI as a co-pilot can be extraordinarily powerful, especially in complex tasks where humans experience fatigue. In most cases, an individual radiologist with enough time is still better than the AI-driven results alone. For example, the article mentions a 2021 review that found that in 34 of 36 studies, the AI did worse than a single radiologist screening for breast cancer. However, the outcome from humans working with AI is modestly better and takes much less energy.
All of this points to the need for human-centered AI. We can all take inspiration from this example to design systems that are intentionally meant to augment decisions and that expert humans feel confident about trusting.
What it's about: Conversations surrounding ChatGPT have exploded over the last few months. If you’re unfamiliar with the model, you’ll find this user guide from Allie K. Miller helpful.
Unlike familiar machine learning models like Siri and Alexa, ChatGPT has a full conversational interface. The ChatGPT model differs because of its chat interface, early guardrails, and context length. In just five days, ChatGPT reached one million users, which is a testament to how easily the public can use its no-code interface. Upon first opening the application, users can see how the team behind ChatGPT is invested in safety mitigations and gathering feedback.
Perhaps the most exciting development seen in ChatGPT is the increased context length and its implications. Context length as it relates to ML models is an indicator of the system’s long-term memory and knowledge retention capabilities. ChatGPT has a context length of at least four times greater than GPT-2, which was released in February 2019 by OpenAI. 100 tokens equal about 75 words, so ChatGPT is theorized to be able to handle between 3,072 to 6,144 words.
For professionals, ChatGPT is a tool that can be leveraged to improve work life, but it can’t do everything. In the simplest terms, the system is an augmentation that relies on your experience, execution, critical thinking, and emotional intelligence. A meaningful way to use ChatGPT could be to let it handle your repetitive tasks at work (see Allie’s guide for specific examples). Don’t forget that ChatGPT is a research tool before anything else, and it’s important not to submit confidential information.
Why it matters (from Nicholas Napier): A big win with ChatGPT is the ease of interaction and turning of the output. Deploying an API that enables non-experts to interact helps demystify that incorporating these tools into daily life doesn't need to require an application to difficult real-world problems.
The second is being able to adjust the complexity of the output. Although context prompting was relevant in GPT3, the ability to interact with the system and modify the output iteratively is a huge step in fine-tuning talks, papers, and social media posts to a wider audience.
While these additions are great, we are also exposing a wider audience to the fact that AI/ML is not perfect. I would almost argue that the errors ChatGPT is producing are going to drive the movement of improved trust in AI, as everyone can see that these systems aren't perfect and human interaction is still needed. We aren't trying to take jobs with these tools, but alleviate most of the effort behind mundane but important tasks.
How Foundation Models Can Advance AI in Healthcare
What it's about: Transitioning from tech demos to the AI deployment has been a challenge in the healthcare sector, but a new class of foundation models may lead to more affordable, easily adaptable health AI. Foundation models draw on classic ideas from deep learning by learning from large amounts of unlabeled data and being adaptable with better sample efficiency.
The article explains the opportunities foundation models offer in terms of a better paradigm of doing “AI in healthcare” by outlining what foundation models are and their relevance, and highlighting key opportunities that they provide. Some of these opportunities include:
AI adaptability with fewer manually labeled examples
Modular, reusable, and robust AI
Making multimodality the new normal
New Interfaces for human-AI collaboration
Easing the cost of developing, deploying, and maintaining AI in hospitals
Why it matters: In healthcare, the information within electronic health records (EHRs) can be used to learn classification, prediction, or survival models to assist in diagnosis. The data can also help with proactive intervention. Despite solid predictive performance, models trained on this kind of data do not translate into clinical gains, like better care or lower cost.
This gap is called an AI chasm. In addition to concerns that models are not reliable or fair, the AI chasm is important to understand because it can help us reduce the time and energy spent on training models so we can then focus on creating model-guided care workflows, as well as models that are useful, reliable, and fair.
A big topic surrounding AI concerns the cost of implementation. Learning models can require custom data that cost over $200,000, and end-to-end projects cost over $300,000, so the total cost of ownership of healthcare “models” is too high. The current paradigm is unsustainable. Instead of building single models, we should shift our focus to creating models that are cheaper to build, use reusable parts, can handle multiple data types, and are resilient to changes in data.
Foundation models also offer new interfaces for human-AI collaboration. Opportunities for interacting with AI models include natural language interfaces and the ability to engage in a dialogue. By investing in building collections of natural language instructions, instruction tuning can be used to improve generalization in models. Adopting foundation models in healthcare will put human-AI collaborations front and center.
What it's about: A shortage of trained talent could be slowing down advances in AI by hindering opportunities to scale AI solutions across organizations. Overall, the talent shortage could result in a significant imbalance in AI adoption and scalability.
Rather than lack of skills, the underlying problem is that individuals aren’t connecting with the right opportunities. That is to say, the AI industry is not doing enough to offer the right platform for talented people to launch their careers.
The issue? Many organizations do not have best practices in place for the next wave of deep learning, and others are still in early stages of AI adoption. With companies that do have existing artificial intelligence and machine learning talent, a solid talent development strategy is missing.
The article lays out a solution for creating a strong talent development strategy, consisting of:
Identifying those best fit for enablement programs
Enabling career transitions
Building robust best-in-class in-house learning platforms
Nurturing partnerships with startups, and MOOC platforms
Nurturing partnerships with universities and think tanks
Initiating mentoring programs from experienced AI professionals
Creating incentives
Sponsoring temporary gig projects and job rotation
Instituting hackathons and Ideathons
Creating a steady pipeline of entry-level talent
Creating learning opportunities
Why it matters: In what many experts are referring to as the industrial age of AI, we can’t afford to slow down advances. A talent shortage is affecting almost every industry, but for different reasons.
When it comes to data science talent, or the “builders” of AI, we aren’t not lacking the necessary skills, we just need better platforms for connecting the right people to the right opportunities to grow their careers, and consequently, grow AI implementation overall. Many extraordinary people already exist in the market who would be perfect fits for a career in AI, so what can we do to find them?
As a relatively new industry, those working in AI are still figuring out talent development, which is critical if we want to seize the opportunities provided by the next generation of the AI revolution. Getting a talent pool from relevant backgrounds like mathematics, statistics, computer science, and economics means individuals are already acclimatized to structural problem-solving. While employing those with diverse backgrounds (professionally and culturally) ensures that you’re building equally as representative models.
Additionally, nurturing partnerships plays a huge role in growing AI talent. In other words, we need the support of universities, think tanks, and established professionals. Developing learning platforms and mentoring programs, which exist in other industries, can also help attract and grow the right career professionals. While it’s amazing that AI growth is happening exponentially, it’s also important to get back to the basics to ensure sustainability within the industry.
How will sectors be impacted differently by AI regulation?
With the proliferation of AI across all market sectors, we've already begun to see significant scrutiny by regulatory agencies. We can expect that some sectors will be subject to more significant regulation than others, and this will likely directly correlate with the regulatory frameworks that govern them today.
The EU’s AI Act, which we've discussed before, is currently the largest and most visible piece of AI legislation that has been proposed. The act takes a risk-based approach, scaling requirements based on the potential risk of a particular AI application. Building from this logic, we can expect more heavily regulated industries like healthcare, finance, and transportation to be more significantly affected, as unintended consequences from AI applications in these sectors have the potential to be more significant.
When considering how AI regulation is likely to impact your sector or industry, ask yourself “what are the consequences of getting something wrong?”
This is a fundamental question that should be asked of any AI endeavor, and the answer will give you a good indication of what you can expect to see AI regulations focus on. As the landscape rapidly changes, reviewing guidelines that have been developed by relevant industry associations and advocacy groups like the Responsible AI Institute is a strong starting point. Ensuring your organization is following these guidelines now will help your organization be much more prepared for future regulations.
- Nicole McCaffrey // Chief People & Marketing Officer and Bob Wood // Data Science Consultant, L1
